We introduce a parsimonious framework for choosing among alternative expected-return proxies (ERPs) when estimating treatment effects. By comparing ERPs’ measurement-error variances in the cross-section and in time series, we provide new evidence on the relative performance of firm-level ERPs nominated by recent studies. Generally, “implied-costs-of-capital” metrics perform best in time series; while “characteristic-based” proxies perform best in the cross-section. Factor-based ERPs, even the latest renditions, perform poorly. We revisit four prior studies that use ex-ante ERPs and illustrate how this framework can potentially alter either the sign or the magnitude of prior inferences.